fix(data-plane): log single-controller metrics before the step is committed

[ZhiyuLi-Nvidia]

Fixes a silent metric loss on the async single-controller path, on top of #3616.

The bug

log_metrics(..., step_finished=True) calls run.log(..., commit=True), which
closes that wandb step. wandb accepts a later log against a closed step,
returns normally, and discards it — no exception, no falsy return.

single_controller.py called the data-plane logger immediately after that
commit:

self._logger.log_metrics(timing_metrics, ..., step_finished=True)  # closes step N
self._log_data_plane_metrics(total_time)                           # dropped

So every data_plane/* series on the async path was computed — snapshot
gathered, deltas taken, breakdown table built — and thrown away.

Why it went unnoticed

• Silent by construction. No exception, nothing falsy to check. The only
  symptom is an empty chart.
• stdout still looked healthy. The • data plane: …ms, … MB moved print
  lives inside the discarded call.
• Invisible to unit tests. A fake logger records the call happily; the drop
  happens inside wandb.
• The comment asserted the opposite. # step_finished=True here since this is the final log of our current step stopped being true when a log was added
  after it.

grpo_sync had the identical bug. It was caught only because a real GRPO run
showed 85 logged keys and zero data_plane/*, and was fixed there. This path
was never exercised by a run — its docstring says it mirrors grpo_sync, and
it mirrored this too.

The fix

Move one statement above the commit.

The test

Asserts the invariant for every algorithm module that logs data-plane
metrics, rather than for one call site — a third wiring would otherwise repeat
it. Source order is the only observable, since the drop happens inside wandb.

Two details worth review:

• It compares the call site (rindex), not the first occurrence. The first
  occurrence is the method definition, which trivially precedes everything —
  my first draft of this test passed against the unfixed file.
• Comment lines are stripped before searching, because the comments name
  step_finished=True too.

Verified by running the test's logic against both versions:

	invariant holds
single_controller before fix	✗ (test fails, as intended)
single_controller after fix	✓
grpo_sync	✓

Base

Branched off zhiyul/data_plane_observability_metrics rather than main,
because _log_data_plane_metrics in single_controller.py only exists in
#3616. Against main this would show 51 commits / 66 files; against the
feature branch it is 2 files.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[ZhiyuLi-Nvidia]

/ok to test d6dbd00